Miguel Aroca-Ouellette
Data AnalystStaff ML Engineer | GenAI & Search | Ex-Zendesk | Ex-Google
Career Break
San Francisco, California, United StatesMember since June 24, 2026
About
Staff ML engineer who bridges AI research and production, from filed patents and publications to systems serving millions of users. Background spans Google, Zendesk, and industry AI research across North America and Europe. At Zendesk, I led the technical delivery of the company's #1 priority initiative: RAG-powered AI agents for customer support, now serving millions of users. I also built semantic search ML infrastructure from the ground up, improving help center search relevance at scale. As an engineering manager, I led and grew ML teams through rapid scaling, re-orgs, and a major acquisition integration. Prior to Zendesk, I developed large-scale ML models for Google Ads and conducted AI research at Blue Prism, co-inventing a one-shot learning method with a filed international patent.
Top Skills
Machine LearningPythonResearchC++Microsoft WordPowerPointMicrosoft ExcelTeamwork
Skills
Machine LearningPythonResearchC++Microsoft WordPowerPointMicrosoft ExcelTeamworkPublic SpeakingCustomer ServicePhotoshopSocial MediaMicrosoft OfficeMatlabJavaTeam LeadershipEvent Planning
Education
Caltech
Master of Science (M.S.), Electrical Engineering
January 1, 2015 – June 1, 2016
McGill University
Bachelor of Engineering (B.Eng.), Honours Electrical Engineering
January 1, 2011 – January 1, 2015
Experience
Career Break
Travel
May 1, 2025 – Present
Zendesk
Engineering Manager, Machine Learning
October 1, 2023 – May 1, 2025
Zendesk
Staff Machine Learning Engineer
March 1, 2023 – October 1, 2023
Zendesk
Senior Machine Learning Engineer
November 1, 2021 – March 1, 2023
Blue Prism
Research Engineer
February 1, 2020 – October 1, 2021
London Area, United Kingdom
Projects
Guerilla: A Deep Learning Chess Engine
August 1, 2016 – May 1, 2018
Guerilla is a deep learning chess engine developed by myself and Stephane Aroca-Ouellette. At the moment, the top chess engines in the world rely on heuristic features chosen by expert chess players with carefully tuned weights (their relative importance to each other) in order to evaluate the strength of a given position. The goal of this project was to develop a chess engine which uses a machine learning model, specifically a deep neural network, to replace the expert chess players and allow the engine to select and weigh its own features. In combination with optimized game tree search the engine can thus learn to play high level chess on its own. The code can be found on my GitHub: https://github.com/StephAO/guerilla A detailed, informal introduction to the project can be found on my blog: Part 1: https://unarresteddev.wordpress.com/guerilla-a-chess-engine-part-1/ Part 2: https://unarresteddev.wordpress.com/guerilla-a-chess-engine-part-2/ Post-Mortem: https://unarresteddev.wordpress.com/guerilla-a-chess-engine-post-mortem/
ϵ-Safe: Learning to Predict World Cup Page Requests
June 1, 2016 – June 1, 2016
Predicting the next page given a sequence of page visits from a user one can provide a better user experience through page recommendations, and decreased latency through caching. In this project the HMM model is extended to this online learning problem through a novel learning algorithm which we have named ϵ-Safe. By using an ϵ-greedy bandit to balance between new and old user behaviour information this model achieves significant faster computation times, while maintaining high prediction accuracy. The model was then applied to a World Cup page request dataset, where it achieves better accuracy than a naive predictor and performs almost as well as an offline trained HMM with full-knowledge. The code can be found on my GitHub: https://github.com/miguelarocao/e-safe A detailed, informal explanation of the project and its results can be found on my blog: https://unarresteddev.wordpress.com/e-safe